Use Evolved Sampling 🎭🅐🅓②

If you've built a Gen2 AnimateDiff chain - Load AnimateDiff Model → Apply AnimateDiff Model - and nothing is happening, it's almost certainly because you forgot this node. Everything upstream just prepares data; this is the one that actually patches your model with it and returns something usable in a KSampler. Skip it and your M_MODELS bundle sits there doing nothing.

Here's the part worth knowing even if you're not doing AnimateDiff at all: the name is literal. "Evolved Sampling" - the sliding-context-window sampling technique this whole pack is built around - isn't tied to motion modules. You can leave m_models unplugged entirely and use this node purely to get context-window sampling on an ordinary image workflow, which is exactly what the README means by "usage of Context Options and Sample Settings outside of AnimateDiff."

How it works

This is the patch step. It takes your base model, and optionally an M_MODELS bundle (the accumulated motion models from one or more Apply nodes), context_options (your sliding-window config), and sample_settings (noise scheduling, seed behavior, FreeNoise/FreeInit-style tweaks), and returns a MODEL with all of that baked in as a sampling-time patch. Nothing in the chain before this node has touched your actual model object - this is where it happens.

The inputs and outputs that matter

Required: model (your checkpoint) and beta_schedule (default autoselect - this has to match whichever motion module family you're using if m_models is plugged in; HotshotXL and AnimateDiff-SDXL both need their own non-default schedule per the README).

Optional: m_models - leave empty for context-window-only sampling on a non-AnimateDiff workflow, or plug in the output of an Apply node to actually get motion. context_options for sliding-window length control. sample_settings for the deeper noise/seed tuning bundle from ADE_AnimateDiffSamplingSettings. Output is a single MODEL, wired straight into your KSampler like any other model output.

How to install it

ComfyUI Manager: search AnimateDiff Evolved, author Kosinkadink, install, restart. Manual:

cd ComfyUI/custom_nodes && git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved

If you're using this with motion models, you need at least one downloaded separately (mm_sd_v15_v2 is the safe default; links in the README) placed in ComfyUI/models/animatediff_models. If you're using it purely for context-window sampling with m_models unplugged, no model download is needed at all.

Common issues & troubleshooting

Chain looks right but output shows no motion. Almost always this node missing from the end of the chain, or its m_models input not actually connected to the Apply node's output. Trace the wire, don't just eyeball the graph.

Weird artifacts or broken output when m_models is plugged in. Same root cause as everywhere else in this pack: beta_schedule not matching the loaded motion module's family. Set it explicitly rather than trusting autoselect if you're on HotshotXL or AnimateDiff-SDXL.

Confused about whether you need this if you're not doing AnimateDiff. You don't, ever, unless you specifically want sliding-context sampling for VRAM reasons on a very long or very high-res generation. For ordinary single-image or short-batch work, skip this node entirely and use your normal model output.